A Probabilistic Approach to Detect Urban Regions from Remotely Sensed Images Based on Combination of Local Features Beril Sırmac ¸ek German Aerospace Center (DLR) Remote Sensing Technology Institute Weßling, 82234, Germany E-mail: Beril.Sirmacek@dlr.de Cem ¨ Unsalan Computer Vision Research Laboratory Yeditepe University Istanbul, 34755, Turkey E-mail: unsalan@yeditepe.edu.tr Abstract—Detecting urban regions from very high resolution aerial and satellite images provides very useful results for urban planning, and land use analysis. Since manual detection is very time consuming and prone to errors, automated systems to detection of urban regions from very high resolution aerial and satellite images are needed. Unfortunately, diverse characteristics of urban regions, and uncontrolled appearance of remote sensing images (illumination, viewing angle, etc.) increase difficulty to de- velop automated systems. In order to overcome these difficulties, herein we propose a novel urban region detection method using local features and a probabilistic framework. First, we introduce four different local feature extraction methods. Extracted local feature vectors serve as observations of the probability den- sity function to be estimated. Using a variable kernel density estimation method, we estimate the corresponding probability function. Using modes of the estimated density, as well as other probabilistic properties, we detect urban region boundaries in the image. We also introduce data and decision fusion methods to fuse information coming from different feature extraction methods. Extensive tests on very high resolution grayscale aerial and panchromatic Ikonos satellite images indicate practical usefulness of proposed method to detect urban regions automatically in a robust and fast manner. I. I NTRODUCTION Remotely sensed satellite and aerial images provide very valuable information. However, their covering areas and res- olution make manual analysis very difficult and prone to errors. Furthermore, especially urban areas are dynamic en- vironments. Hence, they should be monitored and analyzed periodically. Due to these problems, developing algorithms to detect particular objects in remotely sensed images is a very important research field. Especially, automatic detection of urban region boundaries can provide useful information to municipalities, mapping agencies, military, government agen- cies, or unmanned aerial vehicle developers. Unfortunately, automatic object detection algorithms cope with some diffi- culties on remotely sensed images. First, buildings generally have diverse characteristics with different texture, color, and shape. In addition to that, in their appearance on the image, the illumination, view angle, scaling, occlusion effects are uncon- trolled. Therefore, classical object detection algorithms cannot provide an acceptable detection performance. Therefore, more advanced methods are required for robust detection of the objects in remotely sensed images. In related literature, due to the importance of the problem, many researchers concentrated on developing automated sys- tems to detect urban regions. Karathanassi et al. [7] used building density information to classify residential regions. They benefit from texture information and segmentation to extract the residential areas. Unfortunately, they had several parameters to be adjusted manually. Benediktsson et al. [1] used mathematical morphological operations to extract struc- tural information to detect the urban region boundaries in satellite images. Their method is based on using a neural network which is trained using training urban area regions to classify input images. ¨ Unsalan and Boyer [16], [18] used structural features to classify urban regions in panchromatic satellite images. Since they use statistical classifiers, they also need training data to detect the urban area in the image. In a following study, ¨ Unsalan and Boyer [17] associated structural features with graph theoretical measures in order to grade the satellite images and extract the residential regions from them. Fonte et al. [5] considered corner detectors to obtain the type of the structure in a satellite image. They concluded that corner detectors might give distinctive information on the type of structure in an image. Bhagavathy and Manjunath [2] used texture motifs for modeling and detecting regions (such as golf parks and harbors) in satellite images. They focused on repeti- tive patterns in the image. Bruzzone and Carlin [3] proposed a context-based system to classify very high resolution satellite images. They used support vector machines fed with a novel feature extractor. Fauvel et al. [4] fused different classifiers to extract and classify urban regions in panchromatic satellite images. Zhong and Wang [20] extracted urban regions in grayscale satellite images using a multiple-classifier approach. These last three studies also need training data for urban area classification. In a related study, Sırmac ¸ek and ¨ Unsalan [16], used scale-invariant feature transform (SIFT) and graph theory to detect urban areas and buildings in grayscale Ikonos images. They used template building images for this purpose. Although graph theoretical methods are suitable for urban